# Machine learning using entropy–based texture features from MRI to differentiate histological subtypes of non–small cell lung cancer identified as metabolically active on PET/MRI

## TL;DR

This study uses MRI and machine learning with entropy-based texture features to better distinguish between two types of non-small cell lung cancer.

## Contribution

The study introduces entropy-based texture analysis to enhance MRI-based classification of NSCLC histological subtypes.

## Key findings

- Combining entropy-based and statistical texture features improved classification performance compared to using statistical features alone.
- Logistic Regression achieved the highest accuracy (0.75) and precision (0.78) after median filtration of MRI images.
- The proposed method shows potential for improving non-invasive diagnosis of NSCLC subtypes.

## Abstract

Texture analysis is a foundational approach in imaging studies and demonstrates excellent diagnostic performance, with radiomic analysis being the most widely used method. New approaches to texture analysis continue to be developed. However, magnetic resonance imaging (MRI)–based radiomics studies for identifying histological subtypes of lung cancer remain scarce. This study aimed to improve the efficiency of the computer–aided non–invasive diagnosis of non–small cell lung cancer (NSCLC) by supplementing the statistical approaches to MRI image texture analysis with entropy–based methods. The study included 31 patients with NSCLC, categorized into two histological groups containing 12 patients (75 images) with adenocarcinoma (ADC) and 19 patients (79 images) with squamous cell carcinoma (SCC). A total of 154 MRI images were annotated using 154 regions of interest (ROIs). ROIs were extracted, filtered using normalize and median filtrations, and analyzed using standard statistical approaches and novel entropy–based methods. Texture features were selected using Select From Model (SFM) protocol and the classified using k–Nearest Neighbors (kNN), Support Vector Machines (SVM), and Logistic Regression (LR), separately. After 5–fold stratified cross–validation, the LR algorithm achieved the highest classification performance (0.75 accuracy and 0.78 presision) on the combined statistical and entropy–based texture features extracted from MRI images after median filtration. The proposed protocol presented higher efficiency than protocols that worked only on the statistical texture features on unfiltered or normalize filtered MRI images; therefore, it may be suggested for further research on the computer–aided diagnosis of NSCLC histological subtypes.